harisang:
This is a nice and clean scheme. Of course, we still have to think about the dynamics of it, as it seems to me similar races to the bottom can happen here; a solver gets reimburses the difference in utility, which, for simple batches where pennying is utilized, will be some value between zero and the pennying amount. Of course, due to competition, the difference again might tend to zero, so you will be losing the pennying amount, so I am not sure how much difference this makes.
I like the suggestion of combining with some randomization. What we have called “bucketing” internally (i.e., putting all solutions with very close obj value in the same bucket and picking one at random, which is quite close to the “add gausian noise” idea of yours) might indeed help to at least reduce the effect of pennying.
Cool, I created a separate topic here not to pollute the discussion here. Also, agree that this bucketing idea would probably work equally well in practice.
1 Like